The financial services sector has many early adopters of sophisticated analytics techniques involving graph computing. Graph AI is regularly used in this industry for applications such as fraud detection, anti-money laundering, and other criminal activities.
Analysts predict significant growth in the consumer lending market, with Allied Market Research anticipating the global fintech lending market to grow at a 27% annual rate over the next eight years. Fintechs are uniquely positioned to target underserved borrowers partly due to their relatively low overhead, which lets them absorb high default rates, and they are pushing traditional lenders to improve their efficiency in order to remain competitive.
At the core of most credit lenders’ business is their credit decisioning model. Applying graph AI, particularly graph neural networks (GNNs), to this challenge has the potential to improve the accuracy of these models, resulting in higher loan values and lower credit expenses. Credit scoring is a form of predictive modeling that typically uses historical data collected about a borrower to determine whether they will repay their loan. Most underserved borrowers lack a long credit history, so companies are looking towards the use of alternate data sources to produce a realistic credit model for these borrowers.
Expanding the number of data sources that lenders use reduces the risk of inaccuracies that could arise in a model based on a single data set and improves their credit modeling performance by providing a more comprehensive view of the borrower. Using multiple data sources typically requires the accommodation of variable data formats, and graphs provide a flexible structure to integrate data from disparate sources. With high volumes of data coming in from silos or different channels, a flexible structure is crucial for enterprises that want to accommodate the constant flood of data and easily bring a new data source online.
Graph neural networks are deep learning frameworks applied to graph-structured data and can be deployed at scale with minimal cost. Historically, most credit scoring systems have relied on traditional statistical techniques such as regression analysis or logistic regression.
Graph neural networks offer an exciting alternative because they can process much larger sets of data than traditional models. Modern techniques involve aggregating neighboring nodes, encoding node properties as features and using these categorical variables to produce a feature vector (an embedding) and feeding it to deep learning layers. This method enables credit lenders to leverage the maximum amount of data possible to capture a holistic view of a consumer’s credit profile with greater speed and accuracy.
The primary concept behind GNNs is the idea of message passing. With each message passing layer, the information gathered about a consumer expands further, capturing more of their credit profile in addition to how they relate to other consumers. Several embeddings can represent a credit profile at a snapshot in time and when changes occur, GNNs can update those embeddings.
A time series-based model may use monthly embeddings to learn long-term temporal aspects of a dynamic graph’s change over time. This type of processing helps determine what patterns of financial habits are predictive of whether a loan will default. For example, if a borrower takes out new lines of credit immediately prior to the time of application, they would be identified as probably having poor credit.
GNNs excel at encoding high-dimensional data, enabling the machine to capture complex, shifting relationships over time and identify nuances within massive, disparate data. These techniques enable enterprises to learn which features and consumer relationships are most important for credit decisioning.
For regulatory reasons, having a methodology for explainability is a critical component of credit decisioning and there are many ways an enterprise could go about ensuring that a loan acceptance or denial is valid. Having measures in place to track features that propagate through the neural network also allows enterprises to identify and remove which features are producing a bias in the model in order to provide a non-discriminatory credit decisioning model. In some instances, assigning importance scores to embeddings that can be backed into the raw features (such as payment status on an existing tradeline) is an efficient way to identify why a loan gets approved or not.
Using alternative data sources and augmented analytics, financial institutions can explore the untapped credit signals that exist in complex relationships. They can do this by leveraging the maximum amount of data to drive profitability in a growing consumer lender market. GNNs provide a method to determine similarities among consumer, tradeline, and institutional relationships.
Katana Graph brings your data into brilliant focus faster than ever before — without disruption, overspending, or overhaul. To learn more about the Katana Graph Intelligence Platform speak to our graph expert.
Get to Know Katana Graph
We thrive on testing new and diverse ideas.
Katana Graph was born of cutting-edge research and scientific rigor, and these beginnings have a powerful effect on who we are to this day. We’re devoted to problem solving, and are relentless in our pursuit of more effective and more efficient solutions to real-world challenges. Continuous improvement is the very foundation of our success.
